Start a new, empty workflow draft and return its workflow_id. Every other tool takes that id. Nothing is persisted beyond the host's store, and creating a workflow never runs anything.

List every workflow draft this host is holding, with node and edge counts. Use it to recover a workflow_id you did not keep, before creating a duplicate.

Return one workflow's full graph — every node with its kind, config and parentId, and every edge. This is the authoring view; use export_workflow for the portable WorkflowSchema document.

Remove a workflow draft from this host's store. Irreversible on a store that does not version, and it deletes only the draft — nothing that was already exported or run.

Add a node of a given kind to a workflow. The kind is checked against the live registry (call list_node_kinds first — a host may have registered its own). Omitting config applies the kind's schema defaults. Pass parent_id to place the node inside a lane. Returns the created node.

Remove a node and every edge touching it. The edges go WITH the node deliberately — leaving them would produce edges pointing at nothing, reported as a second error the author did not cause.

Merge config into a node, keeping the kind's other defaults. A config that does not satisfy the kind's schema is WRITTEN ANYWAY and reported as warnings — a half-configured node is a normal intermediate state when building a graph one step at a time, and refusing the write makes it impossible to get there. validate_workflow is where 'not finished' is said.

Draw an edge from one node's output port to another's input port. Omit the port names to use each kind's defaults. Both nodes must already exist — connecting to a node that is not there is an error rather than a promise to create it.

Remove one edge, named either by its edge id or by the source/target pair it connects. Removing an edge never removes the nodes it joined.

Every node kind this host has registered, read from the LIVE registry rather than a fixed list — so kinds the host added itself, and kinds vendored in from the marketplace, are included. Each says whether this host ADMITS it, and a refusal names the missing capability.

One kind in full — its config schema, its default config, and its input and output ports. Call this before configure_node rather than guessing field names; the fields are deliberately not on list_node_kinds, which would turn a vocabulary query into a payload nobody reads.

Check a workflow and return every issue TAGGED by who refused it: source "schema" means the graph is malformed and editing it is the fix; source "host" means this host will not run that kind and the fix is a capability or a different kind. Collapsing the two costs you the wrong fix.

Emit the portable WorkflowSchema document — the same JSON the TypeScript, PHP and Python runtimes all read. This is what you hand to another runtime, commit to a repo, or pass to import_workflow.

Read a WorkflowSchema document into a NEW draft and return its workflow_id, plus any issues found. Import is lenient on purpose: a document referencing a kind this host has not registered still imports, with the problem reported — refusing would lose the other forty nodes to fix one.

Execute the graph and return per-node outputs plus ok/error. This is a SMOKE TEST of wiring and routing, not a production run: it takes no executors, so nodes fall back to their kind's own behaviour and anything needing a host capability (a terminal session, an LLM client) refuses rather than acting. Refuses outright to run a graph with validation errors, because a malformed graph produces failures that look like engine bugs and are not. Seed entry nodes with initial_inputs, keyed by node id then port.